| Software agents are software entities situated within a real computational environment. They sense from that environment by their sensors and act on that environment with their effectors. Their actions can affect what they sense in the future, and are in the service of their own drives. This dissertation is concerned with natural language sensing and metacognition in such software agents.;Software agents usually need to communicate with humans in natural language. The design of natural language sensing systems is often a success factor in producing software agents that must communicate with humans. In this dissertation, a detailed description of an implemented natural language sensing system for software agents "living" in a narrow domain is provided. This is presented in the context of Virtual Mattie (VMattie), an autonomous clerical agent who "lives" in a Unix system, automatically reads and understands email concerning seminars (in natural language), and composes and distributes weekly seminar schedule announcements. VMattie's natural language sensing system combines statistical corpus-based approach, surface feature analysis, and information extraction technique. The mechanism used to implement the system is a Copycat-like architecture, with major changes to fit the specific domain.;Metacognition includes metacognitive knowledge, metacognitive monitoring, and metacognitive regulation. The modeling of metacognition in software agents is a new frontier. Building metacognition in agents allows them to overcome internal disorders, to choose an efficient strategy, to optimize resource allocation, and to exhibit human-like behaviors. This dissertation discusses the theory of metacognition, Minsky's B-brain, and Sloman's meta-management. This dissertation explores different computational models, and proposes three general architectures along with their underlying suitable mechanisms for building metacognition in software agents. These general architectures will help other researchers build metacognition in their own agents.;"Conscious" Mattie (CMattie), the successor of VMattie, is a human-like agent in that she implements human cognitive features such as learning, metacognition, emotion, episodic memory, and even "consciousness" in the sense of implementing Baars' global workspace theory of consciousness. This dissertation focuses on the design and implementation of metacognition in CMattie using three different computational models, the classifier system, the fuzzy controller, and the fuzzy classifier system. |
